LineSegNet: A network architecture for extracting continuous lines from complex backgrounds

Line extraction is a key step in the analysis and identification of line chart in paper documents. However, challenges such as noise interference, line discontinuity, and complex image backgrounds often degrade the accuracy and effectiveness of line extraction tasks. In this paper, an end-to-end multi-path lightweight nebvork architecture is proposed, which called LineSegNet for Line extraction in paper documents. LineSegNet are generated in three modules: global perception module, local perception module and progressive fusion module. With these modules, the multi-layer residual structure incorporating attention mechanism and the multi convolutional kernel are integrated in a unified architecture. Global perceptual features and local perceptual features are fused in the progressive fusion module to accurately segment lines from complex backgrounds. The experimental results indicate that our method can significantly reduce hardware resources and demonstrate competitive performance.